PulseAugur / Brief
EN
LIVE 22:07:53

Brief

last 24h
[2/2] 222 sources

Multi-source AI news clustered, deduplicated, and scored 0–100 across authority, cluster strength, headline signal, and time decay.

  1. RESEARCH · 雷峰网 (Leiphone) 中文(ZH) ·

    A palm-sized, 300-gram AI host, why can it run a 122B model?

    Lenovo has launched the P7, a compact AI host weighing 300 grams and consuming 30W, capable of running 122B parameter models locally. This device is designed as an "Agent Computer" for the AI 2.0 era, focusing on continuous, low-power operation for complex tasks. The P7 utilizes a novel computing-in-memory architecture from Post-Silicon Intelligence, specifically the M50 dNPU, to achieve high performance with reduced power consumption and noise. AI

    A palm-sized, 300-gram AI host, why can it run a 122B model?

    IMPACT This new class of compact, low-power AI hosts could enable widespread local inference of large models, driving the adoption of AI agents in consumer and enterprise devices.

  2. RESEARCH · arXiv cs.AI Română(RO) · · [2 sources]

    Real-Time Parallel Counterfactual Regret Minimization

    Two new arXiv papers explore advancements in regret minimization for online optimization and game theory. The first paper introduces a simpler, computationally efficient algorithm for minimizing linear swap regret with a near-optimal bound, leveraging response-based approachability. The second paper presents Parallel CFR, a framework for real-time, depth-limited counterfactual regret minimization that achieves significant speedups by parallelizing iterations and offloading leaf node evaluation to GPUs. AI

    Real-Time Parallel Counterfactual Regret Minimization

    IMPACT These papers advance theoretical and practical approaches to regret minimization, crucial for developing more robust and efficient AI agents in complex decision-making environments.